Improve speed

[thewh1teagle]

Currently kokoro-onnx inference speed is ~2s for each generation on macOS M1 and it uses only ~30% CPU / GPU (CoreML / CPU providers)

Things we tried:

• Using more threads
• PARALLEL mode

Related:

• [Performance] kokoro onnx performance issues microsoft/onnxruntime#23384
• [Performance] ONNX Runtime doesn't parallelize operations in CPU models microsoft/onnxruntime#16158
• the GPU memory usage sometimes spikes unexpectedly #37